CycleGAN for Synthetic Image Generation¶
Collecting data can be expensive and in certain use cases dangerous to acquire. One such example is smoke data. There are many applications relevant to smoke detection with a camera. To skip the data collection stage we create a synthetic dataset of smoky images..
The Full repository for generation and classification is available here: smoke-cyclegan-repository
Synthetic data¶
Synthetic data is information that’s artificially manufactured rather than generated by real-world events. Synthetic data is created algorithmically, and it is used as a stand-in for test datasets of production or operational data, to validate mathematical models and, increasingly, to train machine learning models.
There are a few different ways to generate synthetic data. One possibility is to generate a full new image based on a set of examples. In this post we are exploring image alterations NOT full generation.
CycleGANs¶
To alter the images, we use Cycle Generative Adversarial Network, or CycleGAN. This is an approach to training a deep convolutional neural network for image-to-image translation tasks. The Network learns mapping between input and output images using unpaired dataset.
The common example given is one in which we convert horses to zebras and vice versa
Classifying smoky images¶
The point of generating smoke synthetic images is to avoid having to collect them in real life. This can be dangerous and expensive. Once we have the technology to add smoke to images, we use the images to train a classifier. If the synthetic data truly reflects the real world data, this classifier should be able to detect smoke in the real world. We use a pretrained Resnet-50 model is used to classify the smoky images from the non-smoky ones. It can be found in the Fire-Smoke-Detection folder, adapted from smoke-detection-repository.
To evaluate performance, we convert 500 food images with a camera above a stovetop and split 80/20 training and validation. Our validation performance scores ~96%.